SupportPilot AI Helpdesk

What it does & who it's for

Most AI support tools fail the same way: confidently, about money. SupportPilot is built around the opposite posture — the model drafts only from passages retrieved out of your own documentation, every reply's citations are checked by code before anything is stored or sent, and the four conversations that can actually hurt you — refunds, legal, security, cancellations — are hard-coded to route to a human before a single token is spent. That list is code, not configuration, so no config mistake can turn it off. Everything runs on your infrastructure. Ingestion turns your Markdown, HTML, text and PDF docs into a heading-aware index where the headings become the citations customers see. Retrieval is deterministic BM25 with query-side synonyms and a grounding signal — no embedding calls, no vector database, testable in CI and evaluable offline with no API key (docs/architecture.md defends this choice properly). The model — Anthropic or OpenAI, via plain fetch — writes the final reply under a contract the parser then enforces: no citations, or citations to passages that were never provided, and the draft is rejected and escalated with the rejected draft attached for the reviewing human. An uncited answer never reaches a customer, even when it happens to be right. Two channels ship wired. The chat widget is one script tag, answers with source chips, and hands off politely the moment retrieval can't ground a question. The email channel polls your support mailbox over IMAP and has the feature this category actually needs: shadow mode, where every draft waits in the review inbox and NOTHING sends until a human clicks approve — the going-live checklist walks you from shadow to live in earned stages, and rollback is a one-line env change. Escalations email your team with a deep link straight to the ticket. The evaluation harness is the quality gate: 40 scored sample tickets — messy, typo-ridden, realistic — run against the included sample company (Northbeam Gear Co., 12 docs). Offline, with no key, it verifies deterministically that guarded topics route to a human, answerable questions clear the confidence gate and retrieve the right document, and out-of-scope questions escalate. The shipped configuration passes 40/40, and the demo's terminal replay shows the actual run from when the demo was generated. Rewrite the tickets for your business and the harness becomes your regression suite; the escalation playbook documents the weekly loop that turns escalations into knowledge-base fixes. Before listing, the whole product was exercised on a live boot: real widget conversations (answered with citations, refund guarded, kayaks escalated), a real email run through the parser and pipeline into a cited draft, and the inbox screenshots in this listing are that run — not mockups. A keyless demo provider (clearly watermarked, extractive, double-blocked from ever emailing a customer) lets you evaluate the entire pipeline the same way before adding an API key.

Version history

1. v1.0Latest3 Aug 2026

Licence & terms

These terms are snapshotted at purchase — what you see here is what applies to your order.

Reviews